The Ghost in the Machine: How Archivists Are Using LLMs to Read History's Handwriting at Scale

The problem has haunted historians, genealogists, and archivists for generations: centuries of handwritten documents, diaries, census records, and correspondence sit in vaults and digital repositories, largely inaccessible to anyone who cannot decipher the looping, cramped, and often idiosyncratic scripts of the past. For decades, optical character recognition (OCR) has been the workhorse of digitization, but it was built for the clean, uniform typefaces of printed text. Handwriting remained a stubborn frontier—until now. A quiet but profound shift is underway in archival science, driven by the same large language models (LLMs) that power chatbots and code generators. Archivists are turning to LLMs to decipher handwriting at scale, and the implications for historical research, data accessibility, and how we preserve the past are nothing short of remarkable [1].

The core technology enabling this transformation is not a single product but a convergence of advances in machine learning—specifically, the application of transformer-based architectures to handwritten text recognition (HTR). For years, the field relied on specialized models trained on narrow datasets: a particular scribe's hand, a specific century's script, or a single language. These models worked, but they were brittle. An LLM, by contrast, brings vastly larger contextual understanding to the task. It doesn't just recognize individual characters; it predicts entire sequences of words based on linguistic probability, semantic meaning, and even the stylistic quirks of a particular era. This fundamentally different approach yields results that would have been unthinkable just five years ago [1].

The leading platform in this space, Transkribus, has emerged as a focal point for the archival community's embrace of LLMs. Developed through a series of European research projects, Transkribus evolved from a niche academic tool into a full-fledged platform that allows institutions to train custom models on their own collections. The platform's architecture leverages a combination of convolutional neural networks (CNNs) for image feature extraction and transformer-based language models for sequence prediction. This hybrid approach handles the immense variability of handwriting—the slant of a letter, the pressure of a pen, the fading of ink—while using linguistic context to resolve ambiguities. When a model is 80% certain a particular squiggle is an "e" and 20% certain it's an "o," the language model steps in to determine which choice makes sense given the surrounding words. Traditional OCR simply cannot perform this kind of contextual reasoning [1].

The technical details explain why this is happening now. The transformer architecture, first introduced in the landmark 2017 paper "Attention Is All You Need," has been the engine behind the LLM boom. But applying it to handwriting recognition required solving a different set of problems. Unlike printed text, which has relatively fixed geometry, handwriting is a continuous, flowing signal. There are no clear boundaries between characters, and the same letter can look radically different depending on its position in a word. Early HTR systems struggled with this, often requiring manual segmentation of words into individual characters—a labor-intensive process that defeated the purpose of automation. The breakthrough came when researchers realized that the attention mechanism in transformers could learn to attend to relevant parts of an image without explicit segmentation. Through training on thousands of examples, the model learns which visual features correspond to which linguistic units, and it does so in an end-to-end fashion [1].

This is not merely an academic exercise. The practical impact on archival workflows is staggering. Institutions that adopted LLM-based HTR report transcription accuracy rates that rival, and in some cases exceed, human experts for certain types of handwriting. The U.S. National Archives, for example, has experimented with these systems to transcribe Civil War pension files—a collection of over 1.5 million documents written in a dizzying array of 19th-century hands. Early results suggest that the models achieve accuracy above 95% for clear, consistent handwriting, though that number drops for particularly difficult scripts or damaged documents. Even so, the speed is transformative: what would take a team of human transcribers years to complete can now be done in weeks or months [1].

But the story is not just about speed. It is about scale. The sheer volume of undigitized handwritten material worldwide is staggering. The Vatican Secret Archives alone contain over 85 kilometers of shelving. The British Library holds millions of manuscripts. Local historical societies across the globe have attics full of diaries, ledgers, and letters that no one outside a small circle of specialists has ever read. LLM-based HTR offers the first realistic path to making this material searchable, analyzable, and accessible to a global audience. This is not hyperbole; it represents a fundamental shift in the economics of historical research. For the first time, the marginal cost of transcribing a page of handwriting is approaching zero [1].

The implications for scholarship are profound. Historians have long been constrained by the limits of what they can read. A researcher studying 18th-century maritime trade might read the logs of a dozen ships in a summer. With LLM-powered transcription, they could potentially analyze the logs of every ship that entered a major port over a century. This is not just a quantitative change; it is a qualitative one. It enables new kinds of questions—questions about patterns, networks, and long-term trends that were previously invisible. The ability to perform text mining and natural language processing on historical handwriting opens up entire new subfields of computational history [1].

However, the technology has its critics and risks. The most immediate concern is accuracy. While LLMs are remarkably good at deciphering handwriting, they are not perfect, and their errors are not random. They tend to make mistakes that are linguistically plausible—substituting a word that makes sense in context for the word that was actually written. This is a feature of the language model's design, but it is also a bug. A historian relying on machine-transcribed text might miss a crucial detail because the model "corrected" an unusual spelling or a rare name to something more common. The archival community is acutely aware of this problem. Best practices are emerging that involve human-in-the-loop validation, where the model flags uncertain transcriptions for manual review, and confidence scoring, where users can see how certain the model is about each word [1].

There is also a deeper, more philosophical concern about what is lost in transcription. Handwriting is not just a carrier of information; it is an artifact in its own right. The pressure of the pen, the flow of the ink, the crossings-out and corrections—these are all part of the historical record. A clean, machine-readable transcription strips away this materiality. It flattens the document into pure text, losing the embodied experience of reading a handwritten page. Some archivists argue that this is a necessary trade-off for accessibility, while others worry that it represents a form of epistemic violence, reducing the rich complexity of historical documents to a sterile data set. The debate is ongoing, and there is no easy resolution [1].

The business and strategic implications of this technology are also worth examining. The market for archival digitization services is large and growing, driven by government mandates, academic research, and the booming genealogy industry. Companies like Ancestry.com and FamilySearch have already invested heavily in traditional OCR and manual transcription. The emergence of LLM-based HTR threatens to disrupt this market by dramatically lowering the cost and time required for transcription. This could democratize access to historical records, but it could also concentrate power in the hands of the institutions and companies that control the best models and the largest training datasets. The strategic stakes are clear: whoever masters handwriting recognition at scale will have a significant advantage in the multi-billion-dollar market for historical data [1].

The broader context of this development is the ongoing integration of AI into every aspect of knowledge work. The same week that the IEEE Spectrum article on archival LLMs was published, TechCrunch reported on a new open-source gadget called Clawdmeter, which turns Claude Code usage stats into a tiny desktop dashboard for AI coding power users [2]. This small, niche product is emblematic of a larger trend: the granular measurement and optimization of AI usage. Just as archivists use LLMs to extract meaning from historical handwriting, developers use tools like Clawdmeter to extract insights from their own AI interactions. The underlying pattern is the same: AI is becoming an infrastructure layer that we instrument, measure, and optimize [2].

Meanwhile, in a completely different domain, The Verge reported on the return of the crypto Clarity Act to the Senate, noting that banks are already trying to kill it [3]. This legislative battle, while seemingly unrelated to archival science, is part of the same macro story: the struggle to define the rules of the digital economy. The Clarity Act would provide a regulatory framework for cryptocurrencies, potentially unlocking massive institutional investment. The banks' opposition is a classic example of incumbent resistance to disruptive technology. The parallel to the archival world is clear: established institutions—whether banks or traditional historical societies—often resist technologies that threaten their existing business models or professional practices. The archivists who embrace LLMs are, in a sense, the crypto innovators of their field, pushing against a status quo that has been in place for centuries [3].

The same week also saw the launch of AI IQ, a new site that scores frontier AI models on the human IQ scale, with results that are already dividing the tech community [4]. The project assigns estimated intelligence quotients to more than 50 of the world's most powerful language models, plotting them on a standard bell curve. The creator of the site described it as "super useful" [4]. This development is directly relevant to the archival story because it highlights the contested nature of AI evaluation. If we cannot agree on how to measure the intelligence of a general-purpose LLM, how do we evaluate the performance of a specialized model for handwriting recognition? The archival community is grappling with this question right now. Different models perform differently on different types of handwriting, and there is no standardized benchmark. The field is still in its early stages, and the metrics are still being developed [4].

The convergence of these stories—archival LLMs, developer tools, crypto regulation, and AI IQ testing—paints a picture of an industry in rapid, chaotic transformation. The underlying technology is advancing faster than the institutions and norms that govern its use. For archivists, this is both an opportunity and a challenge. The opportunity is to unlock the vast, handwritten record of human history. The challenge is to do so in a way that is accurate, ethical, and sustainable [1][4].

One of the most interesting aspects of the archival LLM story is the role of the open-source community. Transkribus, while not entirely open-source, has strong roots in the academic research community and has released tools and models that others can use and modify. This stands in contrast to the proprietary, walled-garden approach of many commercial AI companies. The open-source ethos is particularly important for archival work because it allows institutions to train models on their own data, without sending sensitive historical documents to a third-party cloud service. This is a non-trivial concern: many archives contain personal information, medical records, or culturally sensitive materials that cannot be shared freely. The ability to run models locally or on private infrastructure is a key requirement for many institutions [1].

The technical architecture of these systems is also worth examining in more detail. The typical workflow involves several stages. First, the handwritten document is scanned at high resolution, often 300 DPI or higher. The image is then preprocessed to correct for skew, lighting, and other artifacts. Next, a segmentation model identifies lines of text—a non-trivial problem for documents with irregular layouts or marginalia. Each line is then fed into the HTR model, which outputs a sequence of characters. Finally, a language model post-processes the output to correct errors and improve fluency. The entire pipeline is typically implemented as a single, integrated system, but each stage can be tuned independently. The best results come from fine-tuning the language model on domain-specific text—for example, training it on 19th-century correspondence before using it to transcribe 19th-century letters [1].

The training data itself is a critical resource. Building a high-quality HTR model requires thousands of transcribed pages, ideally from the same hand or the same historical period. This creates a chicken-and-egg problem: to transcribe documents at scale, you need a good model, but to build a good model, you need transcribed documents. The archival community has addressed this through collaborative data sharing. Projects like READ-COOP, which coordinates the development of Transkribus, have created shared repositories of training data that any institution can use. This collective approach has been essential to the technology's progress, and it stands as a model for how AI development can serve the public interest [1].

Looking ahead, the trajectory is clear. LLM-based handwriting recognition will become faster, cheaper, and more accurate. The barriers to entry will continue to fall. Within a few years, it is plausible that any historian, genealogist, or curious amateur will upload a photo of a handwritten document and receive a high-quality transcription in seconds. This will fundamentally change the relationship between the public and the historical record. It will also create new challenges around authenticity, provenance, and the interpretation of machine-generated text. The archival profession will need to develop new standards and best practices for working with AI-transcribed materials [1].

The hidden risk that the mainstream media is missing is the potential for a "transcription monoculture." If the entire field converges on a small number of dominant models, trained on a limited set of data, the transcriptions could reflect the biases and blind spots of those models. A model trained primarily on Western European handwriting from the 18th and 19th centuries will perform poorly on Arabic script, Chinese calligraphy, or the cursive of a 20th-century Indian clerk. The technology could inadvertently reinforce a Eurocentric view of history, simply because the training data is skewed. The archival community is aware of this risk, but economic incentives push toward consolidation, not diversity. The challenge will be to build a truly global infrastructure for handwriting recognition that works for all scripts, all languages, and all historical periods [1].

In the end, the story of archivists turning to LLMs is a story about the relationship between technology and memory. We are building machines that can read the past, and in doing so, we are changing what the past means. Every transcription is an interpretation, and every interpretation carries the assumptions of its creators. The great promise of LLM-based HTR is that it can make the past more accessible. The great risk is that it will make the past more uniform, more predictable, and more like us. The archivists leading this work understand this tension. They are not naive techno-optimists. They are grappling with the ethical and practical implications of their tools, even as they push the boundaries of what is possible. The future of historical research will be shaped by how well they navigate this balance between access and authenticity, between scale and care. The machines are reading the past. It is up to us to decide what they will find.

---

References

[1] Editorial_board — Original article — https://spectrum.ieee.org/ai-handwriting-transcription-transkribus-lecun

[2] TechCrunch — Clawdmeter turns your Claude Code usage stats into a tiny desktop dashboard — https://techcrunch.com/2026/05/14/clawdmeter-turns-your-claude-code-usage-stats-into-a-tiny-desktop-dashboard/

[3] The Verge — The crypto Clarity Act returns to the Senate this week. The banks are already trying to kill it. — https://www.theverge.com/column/929752/the-crypto-clarity-act-returns-to-the-senate-this-week-the-banks-are-already-trying-to-kill-it

[4] VentureBeat — AI IQ is here: a new site scores frontier AI models on the human IQ scale. The results are already dividing tech. — https://venturebeat.com/technology/ai-iq-is-here-a-new-site-scores-frontier-ai-models-on-the-human-iq-scale-the-results-are-already-dividing-tech